A research paper out of arXiv argues that showing language models what not to do is just as useful as showing them what to do.
Most in-context learning setups for information extraction — tasks like named entity recognition and relation extraction — feed the model a handful of correct examples and hope it generalizes. The paper behind LC-ICL takes a different approach: it bundles those positive examples with deliberately wrong ones, then attaches error-cause labels explaining why each bad example fails. The idea is that a model told "this prediction failed because it confused a company name with a location" can avoid that specific mistake at inference time. Experiments across multiple datasets show LC-ICL outperforming standard few-shot methods on both entity and relation extraction tasks.
The broader implication is that negative examples carry signal that researchers have largely left on the table. If the gains hold across model families and domains, it could shift how practitioners assemble few-shot prompts — a decision that today is often made by intuition rather than method.
It is worth noting that "outperforms previous methods" is a claim that lives and dies by dataset selection; the paper has not yet passed peer review, and real-world IE pipelines are messier than benchmarks suggest.